A semantic search result is useful only if it also meets the user’s constraints. Language, access scope, date and product filters can remove most of a collection, making a graph that works well for ordinary similarity search much harder to traverse. Faster search that misses the remaining valid records is not a useful optimization.
Weaviate’s October 14 technical account introduces PathSeer support in version 1.40, following research by Zhiyue Li and colleagues. The implementation mixes filter-aware exploration decisions within an existing HNSW graph. It uses a limited two-hop expansion and preserves some nonmatching nodes as bridges, rather than assuming every node excluded from the result can also be excluded from navigation. Read the primary source.
Filter selectivity changes the trade-off
The source distinguishes the research paper’s implementation from Weaviate’s adaptation. Its benchmarks compare strategies across several collections, filter types and selectivities, with hardware and search settings stated. Those conditions matter because checking a filter and computing a vector distance have different costs, and their balance changes across workloads.
For an existing user, the immediate opportunity is to compare strategies without treating a new index build as the only path. A useful acceptance test should retain exact filtered ground truth, inspect recall as well as queries per second, and include the filters users actually combine. Highly restrictive cases and filters poorly correlated with the semantic query are particularly informative.
Adaptive traversal could reduce the need to choose one global compromise for heterogeneous search traffic. Whether it does so in production depends on workload shape and stable recall. The next evidence is a controlled comparison on an application’s own dynamic collections, including latency tails and resource use.
References
Questions, news tips or corrections? Contact [email protected].
